Reducing floating-point (FP) precision is used to trade the quality degradation of a numerical program’s output for performance. Existing tools implement this optimization in a decoupled way by first generating all mixed-precision code variants and next searching the best-performing one. In this paper, we present a holistic approach called PrecTuner to consolidate the code generator and the autotuner via only one parameter r. Initialized by automatically sampled values, r is first used to generate several code variants in the polyhedral model, combining this optimization with various loop transformations. These code variants are then used to solve a performance model expressed in terms of r, possibly under a quality degradation budget. The best-performing value of r is finally predicted without evaluating all code variants. Experimental results of the PolyBench benchmarks on CPU demonstrate that PrecTuner outperforms LuIs by 3.28x while achieving smaller errors, and we also validate its effectiveness to optimize a real-life large-scale application. In addition, PrecTuner also obtains a mean speedup of 1.81x and 1.52x-1.73x over Pluto on single- and multi-core CPU, respectively, and 1.71x over PPCG on GPU.